Re: [Starpu-devel] hwloc CUDA plugin missing

[Samuel Thibault <samuel.thibault@inria.fr>]

Samuel Thibault, le ven. 15 févr. 2019 09:24:11 +0100, a ecrit:
> but I guess StarPU determined that cores 20-23 were fastest to
> exchange data with the GPUs,

Note: this is done by measurements in
src/core/perfmodel/perfmodel_bus.c. In StarPU 1.2 this used to be
done for each CPU core, in StarPU 1.3 this is done for each NUMA node
only. In both cases this fills the cuda_affinity_matrix to tell which
CPU should be preferred.

Now that hwloc provides the information we could make
_starpu_get_cuda_affinity_vector use that information instead of the
measurement, and similarly for OpenCL devices, "yaka", help welcome :)

Samuel